OpenAI reports frontier AI firms now generate 8.3x more output per user

Frontier enterprise users of OpenAI's models now generate 8.3 times as many output tokens per active user as typical firms, up from a 2.6x gap in January, according to two studies the company released this week. The widening gap, drawn from a sample of more than 10 million messages, marks the clearest indicator yet that AI adoption inside companies is splitting into two distinct tiers.

The findings come in a pair of reports: Enterprise Signals, a practical look at agentic AI across OpenAI's enterprise customer base, and a companion working paper, How Organizations Use AI: Evidence from ChatGPT. Together, they argue that enterprise AI is moving from answering questions to carrying out work. The stake for corporate buyers is whether access to better models alone will translate into productivity gains, or whether organizational plumbing, governance, and workflow redesign still matter more.

What does "frontier firm" actually mean?

OpenAI ranks enterprise customers monthly by output tokens per active user. Frontier firms sit in the top 10% that month. Typical firms fall between the 45th and 55th percentiles. As of June, frontier firms generated 8.3x as many output tokens per active user as typical firms — a threefold increase over the 2.6x gap recorded in January.

The gap shows up across industries and company sizes. Intensive AI use is not confined to technology companies, according to the report. Among U.S. public companies in the companion study, enterprise adopters held more assets, employed more workers, and spent more on R&D than non-adopters. Access to models alone does not appear sufficient to scale adoption. Complementary investments in continuous employee learning, shared workflows, data infrastructure, and governance separate frontier firms from the rest.

How agentic is enterprise AI getting?

As of June, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers. Agentic workflows typically produce more output because they execute longer, multi-step tasks. The 64% figure therefore reflects both how often Codex is used and how much output those tasks generate.

OpenAI draws a sharp line between assistants and agents. The reports describe assistants as helping people think through work; agents as helping them complete it. Products such as ChatGPT Work and Codex can use tools, create files, and produce work for review. Instead of asking AI how to prepare a presentation, a worker can now ask an agent to gather information across sources and draft the presentation.

Which functions are adopting fastest?

Software engineering was the first center of agentic adoption. Since February, weekly active enterprise Codex users have grown across nearly every knowledge-work function the report tracks. The growth rates tell a clear story:

• Legal: 108x
• Sales: 41x
• Recruiting: 41x
• Marketing: 26x
• Engineering: 5x

The figures suggest agentic AI is no longer an engineer-only story. Legal teams, sales teams, and recruiters are driving the steepest growth in active Codex usage this year.

At Virgin Atlantic, the shift is already reshaping how the airline operates. Engineering teams use Codex to refactor legacy code in 30 minutes instead of two weeks, according to the report. Product teams at the airline use ChatGPT Work to compress weeks of competitive research into hours. That output feeds directly into the carrier's five-year digital strategy.

What separates frontier firms on advanced capabilities?

Frontier firms pull ahead on the advanced features that turn a chatbot into a co-worker. Each week, 21% of active users at frontier firms use Plugins, compared with 9% at typical firms. Skills adoption tells a similar story: 19% of weekly active users at frontier firms use them, against just 3% at typical firms.

The internal benchmark at OpenAI itself is far higher. Ninety-five percent of OpenAI employees use Plugins weekly, according to the report. That figure functions as a ceiling for what enterprise adoption could look like in a fully wired-up organization.

Plugins bundle capabilities that help agents complete specific workflows. A sales Plugin, for example, can combine a team's playbook with access to its CRM. The agent then uses current customer information and past proposals to prepare a tailored response for review.

Who inside companies actually uses AI?

The data contradicts a common assumption. Many surveys report higher AI use among leaders and executives. OpenAI's administrative data, drawn from millions of conversations, finds the opposite. Six months after adoption, early-career employees sent 13 more messages per week than executives.

Usage is highest among early-career workers and falls among more senior employees, the working paper reports. The pattern suggests a potential comparative advantage in using AI among those with less tenure.

For leaders, the result points to a practical move: identify employees with the strongest AI habits and make their workflows visible across the organization. The goal is to let effective practices spread beyond the individuals who invented them.

Why does the frontier gap matter?

Companies may have access to the same models, but frontier firms are putting them to work faster and more deeply across their organizations. The practical agenda OpenAI draws from both studies has three parts:

• Connect agents to the context and tools needed to complete valuable work.
• Establish clear permissions, review, and governance.
• Help employees turn effective individual workflows into shared ways of working.

The gap appears alongside greater adoption of capabilities that connect agents to company context, tools, and repeatable workflows. That pattern points to execution rather than access as the binding constraint on enterprise AI.

What's the so-what for enterprise buyers?

The next 12 months will test whether the 8.3x frontier gap narrows or hardens into a permanent split. OpenAI's recommendation to its enterprise customers is to extend agentic workflows beyond engineering and convert successful individual use cases into repeatable practices. The reports frame that as a closing exercise rather than a moonshot — a matter of permissions, governance, and human review, not new model releases.

OpenAI is also offering enterprise customers a customized benchmark against frontier firms, and it points readers to Enterprise Signals for industry- and function-level breakdowns. The underlying message is blunt: the model wars are over inside the enterprise for now. The competition has shifted to how companies wire those models into the work itself.